Re: High Precision Arithmetic Benchmarks

Joe Keane <>
4 Aug 1996 00:32:44 -0400

          From comp.compilers

Related articles
Re: High Precision Arithmetic Benchmarks (1996-07-28)
Re: High Precision Arithmetic Benchmarks (1996-07-31)
Re: High Precision Arithmetic Benchmarks (Terje Mathisen) (1996-07-31)
Re: High Precision Arithmetic Benchmarks (Joe Keane) (1996-08-04)
| List of all articles for this month |

From: Joe Keane <>
Newsgroups: comp.arch,comp.benchmarks,comp.compilers,sci.math.num-analysis
Date: 4 Aug 1996 00:32:44 -0400
Organization: none
Distribution: inet
References: <4svpfb$> <> 96-07-196 96-07-200 96-07-201
Summary: How about this.
Keywords: arithmetic

I didn't see a specific proposal, so i'll toss one up.

We base the test on GNU MP, using the mpq (rational number) interface.

The specific test is to solve a system of 20 linear equations with
rational coefficients. We use simple Gaussian elimination to compute
the LU factorization and back-substitute to get the solution vector.
The matrix has some known type so the results are easy to check.

Given GMP, the code for this test is not much, maybe 100 lines or so.
If there's interest, i'll write that up and post it for people to run.

Of course, we're largely testing GMP on various machines. From what
i've seen the implementation is quite good, probably difficult to beat.
But certainly people may figure out how to improve parts of GMP on
specific machines; that's fair enough and also a generally useful thing.

You may not change the test program. Buying arrays of multiplier chips
from IBM is frowned on and custom VLSI fabrication is right out.

I think that this is an interesting test from many angles.

Joe Keane, amateur mathematician

Post a followup to this message

Return to the comp.compilers page.
Search the comp.compilers archives again.